Researchers working for Apple and from Cornell University quietly pushed an open-source multimodal LLM in October, a research release called "Ferret" that can use regions of images for queries.